Myth one: tokens are like EcoCash bundles—you buy them once and they sit in your wallet
I heard this from a trader in Bulawayo who sells phone accessories on WhatsApp. He thought AI tokens worked like EcoCash bundles: you pay for 5,000 tokens, they live in your account, and you spend them over a month until they run out. When I explained that tokens reset with every single request, he went quiet for a moment, then said, "So I'm paying for the same conversation twice if I ask a follow-up question?"
Not quite, but close enough to matter.
Tokens measure the length of both your prompt and the AI's reply. If you send a 200-word product description and ask Claude Sonnet 4.5 to rewrite it for Facebook, that request might use 400 input tokens and generate 300 output tokens—700 total. If you then paste the result back and say "make it shorter," you've just sent another 300 input tokens plus your instruction, and the model generates 200 more. You're now at 1,200 tokens across two requests. The model doesn't remember the first exchange unless you include that text again, and every time you do, you pay again.
This is why tokens AI cost adds up fast when you treat the chat like a WhatsApp thread. If you're paying in USD and working from Harare with intermittent connectivity, you can't afford to re-send the same context five times because you forgot to save the good version.
The fix: draft your prompt in Notes or Google Docs first. Get everything—background, examples, constraints—into one message, send it once, and save the reply. If you need a revision, paste only the new instruction and the specific part that needs changing, not the entire history. On Kryotta, Gemini Flash Lite costs $0.0001 per thousand input tokens and $0.0004 per thousand output tokens, so a 1,000-token round trip is about half a US cent. That's cheap until you're doing fifty revisions a day because you're treating it like a conversation instead of a task.
Myth two: the context window is your chat history, so everything you've ever said is free
A freight agent in Harare told me he'd been pasting every WhatsApp order from the past week into Claude because he thought the context window meant "all the previous messages are stored and the AI remembers them for free." He'd built up a 9,000-word running document and was copying it into every new request. His USD bill that month was three times what he expected.
The context window is the maximum amount of text the model can hold in active memory during one request. For Claude Sonnet 4.5, that's 200,000 tokens—roughly 150,000 words. It's not a free storage locker. It's working memory. Every token inside that window, whether it's new or copied from last week, costs you input-token money every time you send it.
If you paste a 5,000-word order log into every request, you're paying for 5,000 words of input tokens on each request, even if the model only needs the last three orders to answer your question. The window doesn't distinguish between "old stuff you sent before" and "new stuff." It's all input, and input costs money.
The practical difference: if you're asking the model to compare this week's orders to last week's, yes, paste both. If you're just asking it to draft a reply to today's customer, paste only today's conversation. I've seen traders cut their token use by 60% just by trimming the context to what the task actually requires.
One more thing—context window size doesn't mean quality. Gemini Flash has a 1-million-token window, but that doesn't make it better at writing a single product caption than Claude Haiku 4.5, which has a 200,000-token window. The window is a constraint, not a feature. Use it when you need it; ignore it when you don't.
Myth three: hallucinations are bugs the company will fix, so I should wait for the next version
A woman who runs a small pharmacy in Bulawayo asked me when OpenAI would "patch the hallucination problem" because she'd asked GPT-OSS 20B for drug interaction warnings and it had invented a contraindication that doesn't exist. She thought hallucinations were software bugs—errors in the code that the developers would eventually eliminate.
They're not bugs. They're how the models work.
Large language models predict the next most likely word based on patterns in their training data. They don't check a database of facts. They don't know when they're guessing. If the pattern says "this sounds like a medical warning," the model will generate text that sounds like a medical warning, even if the specific claim is false. AI hallucinations aren't a defect to be patched out; they're a side effect of the architecture. Every model hallucinates—Claude, Gemini, Llama, all of them.
The question isn't whether a model hallucinates. It's how often, and whether you've built a workflow that catches it before it reaches a customer.
For the pharmacy owner, I suggested she use the AI to draft the warning, then check it herself against the product insert or the MIMS database she already subscribes to. The model saves her fifteen minutes of typing; she spends three minutes verifying. That's still a win, and it's honest about what the tool can do. If you're using AI to generate anything factual—dosage instructions, tax advice, import tariffs—you verify. If you're using it to rewrite a caption or brainstorm SMS promotions, hallucinations matter less because you'll spot the nonsense before you send it.
DeepSeek R1 is slightly better at showing its reasoning, which makes some hallucinations easier to catch, but it still invents facts when the pattern fits. No model is safe to trust blindly.
Myth four: grounding means the AI checks its facts, so I can trust the answer
This one trips up even people who've read the feature descriptions. A Harare trader selling imported kitchenware asked me whether he should pay extra for "grounded models" because he thought grounding vs fact-checking were the same thing. They're not.
Grounding means the model pulls information from a specific source you provide—a PDF, a webpage, a pasted document—and bases its answer on that source instead of its general training data. If you upload your supplier's price list and ask Claude to find the cost of a 24-piece cutlery set, it will search that document and quote the figure from it. That's grounding. It's useful, and it reduces hallucinations about the content of that document.
But grounding doesn't verify whether the document itself is accurate. If your supplier's price list has a typo—$45 instead of $54—the model will confidently repeat the typo because that's what the source says. Grounding is "answer from this file," not "check whether this file is correct."
Fact-checking requires a second step: you, or another tool, comparing the AI's answer against a known-good reference. If you're generating import duty estimates, you ground the model in the ZIMRA tariff schedule, then you check the final figure against the ZIMRA website or your clearing agent's quote. The model can't do that second step for you.
On Kryotta, you can upload a PDF and ground Claude Sonnet 4.5 or Gemini Pro in it. That's perfect for "What does page seven say about returns?" It's not a replacement for "Is this tariff rate actually correct?"
Myth five: open-weight models are free, so I should use them to save USD
A Bulawayo mechanic who writes service quotes told me he'd switched to Llama 3.3 70B because "open-source is free." He was half right.
Open-weight models like Llama, Mistral, and DeepSeek are free to download and inspect. The weights are public. You can see how they're built. But unless you're running them on your own server, you're still paying someone to host them and serve your requests. On Kryotta, Llama 3.3 70B costs $0.0006 per thousand input tokens and $0.0006 per thousand output tokens. That's cheaper than Claude Sonnet 4.5 ($0.003 input / $0.015 output), but it's not free.
The mechanic thought "open-source" meant "no cost." It actually means "you can see the code and run it yourself if you have the hardware." Most of us don't. A single Llama 3.3 70B instance needs at least 140GB of GPU memory, which is two or three high-end cloud servers running constantly. The hosting and API fees are what you're paying for, not the model itself.
Where open-weight models do save money: they're often cheaper per token than closed models of similar capability, and they're improving fast. If you're writing straightforward product descriptions or SMS campaigns, Llama 3.3 70B or DeepSeek V3 will do the job for a fraction of what Claude Sonnet costs. If you need nuanced reasoning or you're drafting something customer-facing that can't sound robotic, Claude or Gemini Pro are worth the extra cents.
I ran a test with ten product captions for a Harare clothing retailer. Claude Sonnet 4.5 cost $0.09 for the batch and needed no edits. Llama 3.3 70B cost $0.02 and needed two captions rewritten because the tone was flat. The mechanic now uses Llama for internal notes and Claude for quotes that go to customers. His monthly USD AI budget dropped from $14 to $6, and the quality stayed consistent.
Questions people ask
Do I pay for tokens even if the AI gives me a bad answer?
Yes. You're paying for the computation, not the result. If you send a prompt and the model generates 500 tokens of nonsense, you still pay for 500 output tokens. This is why testing your prompt on a cheap model first—Gemini Flash Lite or GPT-OSS 20B—makes sense before you run it through something expensive.
Can I reuse the same prompt without paying again?
You pay every time you send it. If you've got a prompt that works—"Rewrite this product description for WhatsApp in 40 words, friendly tone, include price"—save it in Notes and paste it each time with the new product details. You'll pay for the prompt tokens again, but at least you're not rewriting the instructions from scratch and wasting time.
Which model costs least per thousand tokens when USD is tight?
Gemini Flash Lite at $0.0001 input / $0.0004 output, then GPT-OSS 20B at $0.0002 / $0.0002. Both are fine for simple rewrites and list generation. If you need better reasoning, DeepSeek V3 at $0.00027 / $0.0011 is the next step up and still cheaper than Claude Haiku 4.5.
Does a bigger context window mean I should paste everything I have?
No. A bigger window means you can paste more if the task requires it, not that you should. Paste only what the model needs to answer the question. Everything else is wasted tokens.
I've watched traders in Harare and Bulawayo cut their AI costs by 40–60% just by understanding these five points. The models aren't trying to trick you—they're just tools with specific, sometimes counterintuitive rules. Once you know the rules, you stop paying for confusion. If you want to test these ideas with a few prompts and compare model costs side by side, Kryotta lets you run the same prompt through five models at once and shows you the token count for each. You'll know exactly what you're paying for before the bill arrives.
